RFR(S): 8031968 : Mac OS X: VM starts the agent by calling both Agent_OnAttach and Agent_OnAttach_L functions if its agent library is dynamically linked.

[Daniel D. Daugherty]

 > http://cr.openjdk.java.net/~sla/8031968/webrev.01/

src/os/bsd/vm/os_bsd.cpp
     Good to limit the "RTLD_FIRST" use to MacOS X.

     Perhaps a small comment to explain why MacOS X needs to be
     different than other BSDs (or *NIXes):

     // MacOS X needs to use RTLD_FIRST instead of RTLD_LAZY
     // to avoid finding unexpected symbols on second (or later)
     // loads of a library.

src/os/linux/vm/os_linux.cpp
     No comments.

src/os/posix/vm/os_posix.cpp
     No comments.

src/os/solaris/vm/os_solaris.cpp
     No comments.

Dan


On 1/21/14 7:39 AM, Staffan Larsen wrote:
> Here is a new version with #ifdef in os_bsd.cpp: http://cr.openjdk.java.net/~sla/8031968/webrev.01/
>
> I did experiment with RTLD_FIRST on solaris, but had some problems testing it out. I do think it is more correct, but I also think RTLD_LAZY will work correctly there because we dlopen() agent libraries with RTLD_LOCAL (which makes them invisible when calling dlopen(0, RTLD_LAZY).
>
> /Staffan
>
> On 21 jan 2014, at 14:59, Dmitry Samersoff <dmitry.samersoff at oracle.com> wrote:
>
>> Staffan,
>>
>> I think we should use RTLD_FIRST when it available (i.e. on Solaris
>> too), because it's more correct in this case - we are interesting in
>> symbols from current image only, not from all loaded libraries.
>>
>> But if you have concerns, I'm fain with #ifdef __APPLE__ in os_bsd.cpp
>>
>> -Dmitry
>>
>>
>> On 2014-01-21 17:45, Staffan Larsen wrote:
>>> RTLD_FIRST is also defined on solaris, but we don’t want to use it there (I think).
>>>
>>> /Staffan
>>>
>>> On 21 jan 2014, at 14:36, Dmitry Samersoff <dmitry.samersoff at oracle.com> wrote:
>>>
>>>> Staffan,
>>>>
>>>> It might be better to place
>>>>
>>>> #ifdef RTLD_FIRST
>>>> return (void*)::dlopen(NULL, RTLD_FIRST);
>>>> #else
>>>> return (void*)::dlopen(NULL, RTLD_LAZY);
>>>> #endif
>>>>
>>>> to os_posix.cpp rather than duplicate code because BSDs besides Apple
>>>> use RTLD_LAZY.
>>>>
>>>> -Dmitry
>>>>
>>>>
>>>> On 2014-01-21 17:13, Staffan Larsen wrote:
>>>>> Please review the following fix for how statically linked JVMTI libraries are initialized.
>>>>>
>>>>> The problem is that dlopen() works differently on OS X than on Linux/Solaris. We are using dlopen(NULL, RTLD_LAZY) to get a handle for the main program. On OS X, the call should be dlopen(NULL, RTLD_FIRST).
>>>>>
>>>>> The consequence of this is that on OS X if an agent defines Agent_OnAttach_L but is not statically linked, the JVM will not find that symbol the first time the agent is loaded (which is as expected), but the second time the same agent is loaded, the symbol will already exists in the process and now the JVM will call it (which is not expected).
>>>>>
>>>>> webrev: http://cr.openjdk.java.net/~sla/8031968/webrev.00/
>>>>> bug: https://bugs.openjdk.java.net/browse/JDK-8031968
>>>>>
>>>>> Thanks,
>>>>> /Staffan
>>>>>
>>>>
>>>> -- 
>>>> Dmitry Samersoff
>>>> Oracle Java development team, Saint Petersburg, Russia
>>>> * I would love to change the world, but they won't give me the sources.
>>
>> -- 
>> Dmitry Samersoff
>> Oracle Java development team, Saint Petersburg, Russia
>> * I would love to change the world, but they won't give me the sources.
>